Krea2 Two Character Composer:
Krea2TwoCharacterComposer helps you create a still Krea 2 image with two distinct characters while keeping a shared scene visually coherent. It gives each character its own prompt and optional LoRA, then applies those character-specific instructions to separate image regions. You can arrange the characters side by side, one above the other, in two boxes, or within masks you provide. The node also offers controls for how strongly the regions remain separate during sampling and how smoothly character LoRA effects meet at their boundary. It prepares the regional conditioning and patched model for sampling; it does not generate or decode the finished image itself.
Krea2 Two Character Composer Input Parameters:
model
The Krea 2 MODEL to use for generation. The node checks that the connected model is recognized as Krea 2, then patches it to apply the two regional prompts and character LoRAs during sampling. Connect a Krea 2 model rather than an SD or SDXL model. No numeric range or default is defined; a connection is required.
clip
The CLIP text encoder used to turn your scene and character descriptions into conditioning. Connect the CLIP that belongs with your Krea 2 setup. When a prompt needs NegPiP handling, the node uses a patched encoder for that prompt; otherwise it uses the connected encoder. No default or numeric range is defined; a connection is required.
latent
The LATENT that sets the actual image dimensions and passes through to the sampler. The node accepts a still-image latent with the channel count expected by the connected Krea 2 model; its derived width and height must each be at least 64 pixels and divisible by 16. Preview dimensions do not resize it. No default is defined; a connection is required.
base_prompt
Your shared description of the scene, such as the setting, background, lighting, and style. It applies across the image, and the node adds a placement sentence based on the characters’ regions. Keep character-specific trigger words and details in their respective character fields. This is a required STRING field with an empty-string default and no stated length limit; whole-image LoRA tags are not accepted here.
negative_prompt
Optional text describing things to avoid across the image. A nonempty value requires the NegPiP Prompt pack: the node merges it into the encoded scene prompt as negative-weight conditioning, including when sampling at CFG 1. Without NegPiP, leave this field empty. This is a required STRING field with an empty-string default and no stated length limit; LoRA tags are not accepted here.
character_1_trigger
Optional trigger word or phrase for Character 1, placed first in that character’s regional prompt and kept out of the shared scene prompt. Use it when a selected character LoRA calls for a trigger. It defaults to an empty string and has no stated length limit. You must supply either this trigger or a nonempty character_1_prompt after LoRA tags are removed.
character_1_prompt
The description read by Character 1’s region. Describe appearance, clothing, pose, and other details you want associated with this character. Inline LoRA tags can also be placed here; with NegPiP installed, negative weights such as (beard:-0.6) can apply to this character’s prompt. It defaults to an empty string and has no stated length limit. If its trigger is also empty, the node cannot compose Character 1.
character_1_lora
An optional LoRA selected from the available LoRA files for Character 1’s region. The menu contains a no-LoRA choice followed by files found in ComfyUI’s LoRA list; its default is the no-LoRA choice. Choose a listed file when you want its character-specific effect, and use character_1_strength to set its intensity. The source does not provide the literal no-LoRA menu label.
character_1_strength
The strength of Character 1’s menu-selected LoRA. Its default is 1.0, minimum is -10.0, maximum is 10.0, and step is 0.05. A value of 0 disables that menu selection. This setting does not replace strengths written in inline LoRA tags; if the same LoRA appears more than once for this character, the first selection takes precedence.
character_2_trigger
Optional trigger word or phrase for Character 2, placed first in that character’s regional prompt rather than the shared scene prompt. Use it when a selected character LoRA needs a trigger. It defaults to an empty string and has no stated length limit. You must supply either this trigger or a nonempty character_2_prompt after LoRA tags are removed.
character_2_prompt
The description read by Character 2’s region. Use it for this character’s appearance, clothing, pose, and other individual details; inline LoRA tags may also be placed here. With NegPiP installed, negative weights in this prompt can be applied specifically to Character 2. It defaults to an empty string and has no stated length limit. If its trigger is also empty, the node cannot compose Character 2.
character_2_lora
An optional LoRA selected from the available LoRA files for Character 2’s region. Its options are the no-LoRA choice and the files found in ComfyUI’s LoRA list; the default is the no-LoRA choice. Select a file to add its effect to this character’s area, then adjust character_2_strength as needed. The source does not provide the literal no-LoRA menu label.
character_2_strength
The strength of Character 2’s menu-selected LoRA. Its default is 1.0, minimum is -10.0, maximum is 10.0, and step is 0.05. A value of 0 disables that menu selection. Inline tags can supply additional LoRAs, but if the same file is selected more than once for this character, the first selection wins.
direction
Chooses the initial arrangement of the two character regions: Left / Right, Top / Bottom, or Free boxes. The default is the first entry in the node’s direction list; the provided source does not show that list’s literal ordering. Choose Free boxes to use coordinates from boxes. A connected character mask replaces that character’s calculated area regardless of this layout choice.
character_1_percent
Sets Character 1’s share of a left/right or top/bottom split: the left portion or the top portion, respectively. Character 2 receives the remainder, 100 - character_1_percent. The default is 50, minimum is 1, maximum is 99, and step is 1. Free boxes and connected masks define regions through their own shapes instead.
separation
Controls how the two characters’ prompts interact during sampling. Balanced is the default and preserves scene interaction. Strong also restricts cross-region image attention for the first half of sampling. Soft is experimental and lets the whole image read both character prompts during the last 40% of sampling. These are the three options identified by the node.
boxes
Defines two rectangular regions when direction is Free boxes, with Character 1 first. Enter x, y, width, height; x, y, width, height, using percentages of the image. The field defaults to the node’s DEFAULT_BOXES value, whose literal coordinates are not included in the provided source; an empty value also falls back to that default. Character 1 is in front where boxes overlap. No numeric bounds for individual coordinates are shown here, so use the preview to check placement.
lora_edge_softness
Sets how many pixels the two character LoRA effects take to fade into each other at their boundary; it does not blend their prompts. Use 0 for a hard LoRA boundary, or increase it to soften a visible transition. The default is 0 pixels, minimum is 0, maximum is 256, and UI step is 8.
base_strength
An experimental control for how strongly the characters’ areas follow the shared scene prompt relative to their own prompts. The default is 1.0, which leaves the balance unchanged; lower or higher values alter that balance. The minimum is 0.05, maximum is 2.0, and step is 0.05. Adjust it gradually while comparing results, since the source does not promise a particular visual outcome for each value.
character_1_mask
An optional MASK that replaces Character 1’s calculated region. Connect it when a split or rectangular box cannot describe the area you want. If the MASK contains a batch, only its first mask is used. There is no default mask or numeric range; leaving it disconnected keeps the region calculated from the layout.
character_2_mask
An optional MASK that replaces Character 2’s calculated region. Connect it to place this character in a custom-shaped area; a batch contributes only its first mask. There is no default mask or numeric range. You can supply either character mask independently, with the other character retaining its calculated region.
preview_ratio
Controls only the preview’s aspect ratio before sampling. Auto is the default and uses connected preview dimensions or a recognizable latent resolution; fixed-ratio choices are also available, but their exact labels are not included in the provided source. This setting does not change the generated image dimensions, which come from latent.
width
An optional integer preview width supplied through an input connection. Connect it together with height when you want to specify preview dimensions. It does not resize the latent or set the generation width. No default, minimum, or maximum is stated in the provided source.
height
An optional integer preview height supplied through an input connection. Connect it together with width when you want to specify preview dimensions. It does not resize the latent or set the generation height. No default, minimum, or maximum is stated in the provided source.
Krea2 Two Character Composer Output Parameters:
model
The Krea 2 MODEL patched with the regional behavior needed for this composition. Connect this output directly to your sampler, rather than sampling with the unpatched input model, so the character-specific prompts, LoRAs, and separation settings can take effect.
positive
The regional positive CONDITIONING for the sampler. It contains the encoded shared scene and both character prompts, together with the information the patched model needs to associate each character with its region. It also carries passive metadata describing the prompts and settings; that metadata is not an additional prompt or output socket.
latent
The same LATENT supplied to the node, passed through unchanged. Connect it to your sampler’s latent input. Its dimensions remain the dimensions of the generated image, regardless of preview settings.
region_masks
A MASK batch containing the two regions used for the characters, with Character 1 first and Character 2 second. Use it to preview where the character-specific conditioning applies. The regions reflect a split, free boxes, or any connected replacement masks.
negative
A matching empty CONDITIONING for the sampler’s negative input, used when sampling at CFG greater than 1. It is formed by zeroing the positive conditioning while retaining a compatible structure. Do not interpret it as the text from negative_prompt: that text, when supplied, is handled through NegPiP in the encoded scene prompt.
prompts
A STRING report of what the node encoded and selected. It lists the scene text, both character prompts, each character’s LoRAs and strengths, the layout, separation, LoRA edge softness, base strength, and whether NegPiP was used. It also notes references to undefined common prompts when present. Use it to inspect the composition or diagnose an unexpected result.
Krea2 Two Character Composer Usage Tips:
- Start with a Krea 2 model, its corresponding CLIP, and a compatible still-image latent. Connect the returned
model,positive,negative, andlatentto their respective sampler inputs. - Put lighting, setting, and overall style in
base_prompt. Put each person’s identifying features, outfit, pose, and LoRA trigger in that person’s character fields so those instructions stay associated with the intended region. - Begin with a 50/50 split and
Balancedseparation for a straightforward two-person composition. Adjustcharacter_1_percentto give one character more space, or useTop / Bottomwhen the scene calls for vertical placement. - Use
Free boxesfor deliberate placement and inspectregion_masksbefore sampling. Where boxes overlap, Character 1 takes priority. Connect individual masks when the desired character areas need more precise shapes. - If character traits bleed between regions, compare
StrongwithBalanced. TreatSoftas an experiment for scenes where you want more shared prompt influence near the end of sampling. - Begin with
lora_edge_softnessat 0. Increase it only when the LoRA transition at a region boundary looks too abrupt; this control does not soften the boundary between the character prompts. - Leave
negative_promptempty unless NegPiP is installed. A nonempty negative prompt, or negative-weight syntax in a prompt, invokes NegPiP; ordinary character descriptions do not require it. - Set the image size in the latent-producing node. Use
preview_ratioor the connectedwidthandheightonly to adjust the preview, not to change sampling resolution.
Krea2 Two Character Composer Common Errors and Solutions:
This ComfyUI version has no native Krea 2 support. Update ComfyUI.
- Explanation: The node cannot find ComfyUI’s native Krea 2 model support.
- Solution: Update ComfyUI to a version with native Krea 2 support, then reload the workflow.
Connect a Krea 2 model. The connected model is <model type>.
- Explanation: The connected MODEL is not recognized as Krea 2. - Solution: Connect a Krea 2 model to
modeland use the corresponding CLIP.
Connect a Krea 2 LATENT with <channels> channels, for example from EmptySD3LatentImage. SD/SDXL latents are incompatible.
- Explanation: The latent does not have the channel shape expected by the connected model. For the Krea 2 format described in the supplied code, that is 16 channels.
- Solution: Replace an SD or SDXL latent with a compatible still-image latent, such as one from
EmptySD3LatentImage.
This composer supports still images, not video latents.
- Explanation: The connected latent contains more than one video frame.
- Solution: Supply a still-image latent instead of a video latent.
Krea 2 image dimensions must be multiples of 16, at least 64px.
- Explanation: The width or height derived from the latent is too small or is not divisible by 16. - Solution: Set both dimensions in the latent-producing node to at least 64 pixels and multiples of 16. Preview
widthandheightwill not fix the latent.
Choose Auto or one of the fixed preview ratios.
- Explanation: The supplied
preview_ratiois not one of the node’s accepted choices. - Solution: Select
Autoor a fixed ratio from the node’s menu, particularly when editing an API workflow by hand.
Choose Left / Right, Top / Bottom or Free boxes.
- Explanation: The supplied
directiondoes not match a supported layout choice. - Solution: Select an option from the direction menu or correct the value in the API workflow.
Choose Soft, Balanced or Strong separation.
- Explanation: The supplied
separationvalue is not supported. - Solution: Choose
Soft,Balanced, orStrong.
Base strength must be between 0.05 and 2.
- Explanation:
base_strengthis outside its accepted range or is not a finite number. - Solution: Enter a finite value from 0.05 through 2.0; use 1.0 to return to the unchanged balance.
LoRA edge softness must be between 0 and 256 pixels.
- Explanation:
lora_edge_softnessfalls outside the accepted pixel range. - Solution: Set it from 0 through 256; start at 0 and increase it only if a LoRA boundary needs softening.
Enter a trigger word or a prompt for Character 1.
- Explanation: Character 1 has no regional text after inline LoRA tags are removed.
- Solution: Add a trigger word to
character_1_triggeror a description tocharacter_1_prompt. A LoRA selection alone is not enough.
Enter a trigger word or a prompt for Character 2.
- Explanation: Character 2 has no regional text after inline LoRA tags are removed.
- Solution: Add a trigger word to
character_2_triggeror a description tocharacter_2_prompt. A LoRA selection alone is not enough.
LoRA tags only work in a character's trigger or prompt. For a LoRA on the whole image, use a LoraLoader before the composer.
- Explanation: An inline LoRA tag was found in
base_promptornegative_prompt. - Solution: Move a character-specific tag to that character’s trigger or prompt. For a whole-image LoRA, apply a LoraLoader before this node.
LoRA '<name>' not found. Add it to models/loras and refresh ComfyUI.
- Explanation: A selected LoRA file is no longer present in ComfyUI’s available LoRA list.
- Solution: Check the file name and location under
models/loras, refresh ComfyUI’s file list, and select the available file again.
A NegPiP requirement error when using negative text or negative weights
- Explanation: A nonempty
negative_promptor negative-weight syntax causes the node to require the NegPiP Prompt pack. The provided source does not show the bridge’s exact error wording. - Solution: Install and enable the NegPiP Prompt pack identified by the node, or remove the negative text and negative-weight syntax before running again.
An invalid Free boxes or mask-coverage error
- Explanation: The node parses Free boxes and checks whether the resulting masks adequately cover the image-token grid, but the provided source does not show the exact messages from those helpers.
- Solution: Check that
boxescontains two rectangles in the statedx, y, width, height; x, y, width, heightformat. For connected masks, inspectregion_masksand adjust the mask shapes or layout so the intended areas are covered.
